Feature point selecting system, feature point selecting method and feature point selecting program

1. A feature point selecting system comprising: a recognition task executing unit that executes a recognition task using an importance of each of a plurality of feature point candidates on a three-dimensional shape model for a plurality of evaluation images which are generated from the three-dimensional shape model and which are used to evaluate a recognition error in the recognition task; a recognition error evaluating unit that evaluates a recognition error related to all evaluation images from a difference between a recognition result of the recognition task executing unit and correct data of the recognition task for each evaluation image; a feature point importance determining unit that determines the importance of each feature point candidate by setting a cost function which is a function for the importance of each feature point candidate and which is represented as a function obtained by adding a restriction condition that an importance of an unimportant feature point candidate becomes close to zero, to the recognition error related to the all evaluation images, and calculating the importance of each feature point candidate which minimizes a value of the cost function; and a feature point selecting unit that selects a feature point which needs to be used in the recognition task from the feature point candidates on the three-dimensional shape model based on the importance of each feature point candidate, wherein, with the recognition task executing unit, the recognition error evaluating unit and the feature point importance determining unit, until the value of the cost function which is set based on the importance of each feature point candidate determined by the feature point importance determining unit converges, repeatedly, the recognition task executing unit executes the recognition task, the recognition error evaluating unit evaluates the recognition error related to the all evaluation images and the feature point importance determining unit determines the importance of the feature point candidates.

2. The feature point selecting system according to claim 1 , wherein the feature point importance determining unit excludes from a processing target of the recognition task a feature point candidate comprising an importance equal to or less than a threshold determined in advance.

3. The feature point selecting system according to claim 2 , further comprising: a learning/evaluation data generating unit that creates a plurality of evaluation images from the three-dimensional shape model and correct data of the recognition task for each evaluation image, generates a plurality of learning images which are generated from the three-dimensional shape model and which are used to learn decision data of a feature point extractor, and generates position data representing a position of each feature point candidate on the three-dimensional shape model on each learning image; a local area clipping unit that clips a local area which corresponds to a feature point from each learning image and a local area which does not correspond to a feature point, based on each learning image and the position data generated by the learning/evaluation data generating unit; a feature point extractor learning unit which learns the decision data of the feature point extractor based on the local area which correspond to the feature point and the local area which does not correspond to the feature point; and a feature point extraction executing unit that extracts a feature point from each evaluation image using the decision data.

4. The feature point selecting system according to claim 1 , further comprising: a learning/evaluation data generating unit that creates a plurality of evaluation images from the three-dimensional shape model and correct data of the recognition task for each evaluation image, generates a plurality of learning images which are generated from the three-dimensional shape model and which are used to learn decision data of a feature point extractor, and generates position data representing a position of each feature point candidate on the three-dimensional shape model on each learning image; a local area clipping unit that clips a local area which corresponds to a feature point from each learning image and a local area which does not correspond to a feature point, based on each learning image and the position data generated by the learning/evaluation data generating unit; a feature point extractor learning unit which learns the decision data of the feature point extractor based on the local area which correspond to the feature point and the local area which does not correspond to the feature point; and a feature point extraction executing unit that extracts a feature point from each evaluation image using the decision data.

5. The feature point selecting system according to claim 4 , further comprising a feature point candidate determining unit that determines from the three-dimensional shape model a point which serves as a feature point candidate on the three-dimensional shape model.

6. The feature point selecting system according to claim 5 , wherein the feature point candidate determining unit determines a point which serves as the feature point candidate by applying a plurality of types of feature extraction operators to a texture image of the three-dimensional shape model and a shape of the three-dimensional shape model, and extracting the point which serves as the feature point candidate.

7. A feature point selecting method, comprising: executing a recognition task using an importance of each of a plurality of feature point candidates on a three-dimensional shape model for a plurality of evaluation images which are generated from the three-dimensional shape model and which are used to evaluate a recognition error in the recognition task; evaluating a recognition error related to all evaluation images from a difference between a recognition result of the recognition task and correct data of the recognition task for each evaluation image; determining the importance of each feature point candidate by setting a cost function which is a function for the importance of each feature point candidate and which is represented as a function obtained by adding a restriction condition that an importance of an unimportant feature point candidate becomes close to zero, to the recognition error related to the all evaluation images, and calculating the importance of each feature point candidate which minimizes a value of the cost function; until the value of the cost function which is set based on the importance of each determined feature point candidate converges, repeatedly executing the recognition task, evaluating the recognition error related to the all evaluation images and determining the importance of the feature point candidates; and selecting a feature point which needs to be used in the recognition task from the feature point candidates on the three-dimensional shape model based on the importance of each feature point candidate such that the feature point is selected to match a recognition algorithm in the recognition task.

8. The feature point selecting method according to claim 7 , wherein a feature point candidate comprising an importance equal to or less than a threshold determined in advance is excluded from a processing target of the recognition task.

9. A non-transitory computer readable information recording medium storing a feature point selecting program, which when executed by a processor, performs a method for: executing a recognition task using an importance of each of a plurality of feature point candidates on a three-dimensional shape model for a plurality of evaluation images which are generated from the three-dimensional shape model and which are used to evaluate a recognition error in the recognition task; evaluating a recognition error related to all evaluation images from a difference between a recognition result of the recognition task and correct data of the recognition task for each evaluation image; determining the importance of each feature point candidate by setting a cost function which is a function for the importance of each feature point candidate and which is represented as a function obtained by adding a restriction condition that an importance of an unimportant feature point candidate becomes close to zero, to the recognition error related to the all evaluation images, and calculating the importance of each feature point candidate which minimizes a value of the cost function; until the value of the cost function which is set based on the importance of each determined feature point candidate converges, repeatedly executing the recognition task, evaluating the recognition error related to the all evaluation images and determining the importance of the feature point candidates; and selecting a feature point which needs to be used in the recognition task from the feature point candidates on the three-dimensional shape model based on the importance of each feature point candidate.

10. The non-transitory computer readable information recording medium storing the feature point selecting program according to claim 9 , which when executed by a processor, performs a method for: wherein a feature point candidate comprising an importance equal to or less than a threshold determined in advance is excluded from a processing target of the recognition task.